Advanced Topics in AI

Large-scale foundational Vision-language models (VLMs) are currently transforming computer vision, emerging as a promising path toward robust generalization. This has sparked substantial interest in adapting such foundation models to downstream tasks under realistic constraints: limited labeled data (few-shot adaptation), unlabeled target-domain samples (test-time adaptation), and restricted computational or memory resources (parameter-efficient fine-tuning). In this presentation, I will review recent trends in fine-tuning large VLMs under such practical constraints. I will survey state-of-the-art adaptation strategies and highlight very recent findings that expose important limitations in current evaluation protocols. These results call into question the actual progress suggested by much of the recent literature, particularly approaches relying on convoluted prompt-learning mechanisms. I will emphasize insights grounded in mathematical optimization that, somewhat surprisingly, enable competitive few-shot adaptation, while significantly reducing hyperparameter search and achieving speedups of several orders of magnitude compared to leading methods. I will also present a test-time adaptation approach tailored specifically to VLMs. The talk will be illustrated with several application examples, including use cases in medical imaging.

This lecture presents the theoretical foundations of Graph Neural Networks (GNNs), starting with core models such as Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT) and the principles of message passing and neighborhood aggregation. The lecture covers the common tasks performed on graph-structured data, including node classification, link prediction, and graph classification, along with an analysis of model expressiveness and its limitations. Graph Transformers are also briefly mentioned as an alternative approach that leverages global attention to address some limitations of message-passing architectures.
The lecture then turns to models designed for directed and relational structures, where edge direction and relation type carry essential semantic information that standard GNNs are not equipped to capture. Relational GCNs are introduced as a key architecture for multi-relational data, with particular emphasis on their use to deal with knowledge graphs, where entities and relations play a central role to support tasks such as link prediction.
Finally, the lecture explores extensions addressing scalability and temporal dynamics. GraphSAGE is presented as an inductive framework for generalizing to large-scale graphs while Temporal Graph Networks (TGN) are discussed as a representative approach for capturing dependencies in graphs whose structure and features evolve over time. The goal of the lecture is to provide a critical understanding of how the underlying structure of graph data—relational complexity, scale, temporal dynamics, and long-range dependencies—should guide the selection and design of appropriate GNN architectures.

This hands-on session provides a practical introduction to Graph Neural Networks (GNNs) and their applications across diverse graph learning tasks. We present a unified framework where message-passing architectures—trained through task-specific objectives—consistently achieve competitive performance on challenging benchmarks.
We begin with node classification, demonstrating how Graph Convolutional Networks (GCN) and attention-based mechanisms (GAT) learn node representations by aggregating neighborhood information. We show how architectural choices affect performance on both homophilic and heterophilic networks, and discuss design patterns for building robust node classifiers.
Next, we address link prediction by exploring how GNNs can be extended beyond node-level tasks. We present multiple approaches: heuristic methods, learned similarity metrics, and endto-end architectures. These methods balance computational efficiency with prediction accuracy on standard benchmarks.
For graph classification, we introduce Graph Isomorphism Networks (GIN), graph pooling strategies, and readout mechanisms that aggregate node-level representations into graphlevel predictions. We discuss how architectural choices impact expressiveness and generalization across diverse graph datasets.
Finally, we explore graph generation and clustering, showing how message-passing frameworks enable both discriminative and generative tasks. Throughout, we emphasize practical implementation patterns, scalability considerations, and how to select appropriate architectures for real-world applications.

Modern machine-learning systems are periodically retrained, updated, and replaced to incorporate new data, adapt to distributional shifts, improve accuracy, exploit better architectures. Model updating is a standard practice in deployed AI systems. However, updating a model is not only a matter of improving its average performance. A new model can be globally better than the previous one and still behave worse on specific inputs or with respect to specific trustworthiness properties. This phenomenon is usually perceived by users as a performance regression of the system. In classical software engineering, regression testing is a standard practice to ensure that a system still works as expected after a code change. In machine learning, however, the notion of performance regression is more subtle and is not yet well understood. In this talk, I will discuss regression and backward compatibility in machine-learning systems as an emerging open research problem for trustworthy AI. I will first introduce the general problem, clarifying the relation between regression, non-regression, and backward compatibility in machine-learning model updates. I will then present two research examples. The first concerns security regression, where an updated model for a security task, such as malware detection, fails on threats that were correctly detected before the update, even when the overall detection performance improves. The second example concerns regression of explanations in medical image recognition, where an updated model may preserve or improve predictive accuracy while producing explanations that are inconsistent with those of the previous model.

Modern AI increasingly relies on the ability to perceive, align, and reason across multiple modalities and, more recently, to predict future states of the world rather than merely generate plausible content. This module introduces the foundations of multimodal representation learning, reasoning, and predictive world models: from contrastive vision–language alignment all the way to predictive architectures such as Joint-Embedding Predictive Architecture (JEPA). The theory session covers representation alignment with Contrastive Language-Image Pre-training (CLIP) and its conditional variants, and reasoning with Large Multimodal Models (LMMs) including fine-tuning, adaptation and retrieval-augmented generation. We then dive into predictive world models (vs. generative) and the details of JEPA-like frameworks. In the hands-on lab sessions, we will translate these ideas into code. We will implement and train lightweight CLIP-like models and LMMs, then build a minimal JEPA framework. The final lab block is an open playground in which participants probe key design choices and tinker with the code and models to understand them in depth. By the end of the module, participants will have both a conceptual map of the multimodal-and-world-models landscape and the practical skills to train, fine-tune, and critically analyze such models on their own research problems.

Traditional methods of pattern classification and machine learning usually assume closed world: the input pattern falls within a fixed set of classes. However, in open world, the input pattern can be of either known or unknown classes, or be outlier. While in training, the data may emerge incrementally, and the new dataset contains samples with known or unknown classes, either labeled or unlabeled, or be outlier. Such open-world learning scenario involves multiple challenges including out-of-distribution (OOD) detection, confidence estimation, unlabeled data exploitation, catastrophic forgetting and novel category discovery. The challenges are attacked by combining techniques such as generative modeling, regularization, knowledge distillation, and hybrid learning. This talk will outline the status of open-world pattern recognition, identify the main challenges of open-world learning and main strategies, and present some recent progress achieved in my group: open-set recognition, continual self-supervised learning, class-incremental learning, generalized category discovery, and continual learning for large language model. Finally, I will briefly discuss the status and prospects of continual learning.

Modern AI increasingly relies on the ability to perceive, align, and reason across multiple modalities and, more recently, to predict future states of the world rather than merely generate plausible content. This module introduces the foundations of multimodal representation learning, reasoning, and predictive world models: from contrastive vision–language alignment all the way to predictive architectures such as Joint-Embedding Predictive Architecture (JEPA). The theory session covers representation alignment with Contrastive Language-Image Pre-training (CLIP) and its conditional variants, and reasoning with Large Multimodal Models (LMMs) including fine-tuning, adaptation and retrieval-augmented generation. We then dive into predictive world models (vs. generative) and the details of JEPA-like frameworks. In the hands-on lab sessions, we will translate these ideas into code. We will implement and train lightweight CLIP-like models and LMMs, then build a minimal JEPA framework. The final lab block is an open playground in which participants probe key design choices and tinker with the code and models to understand them in depth. By the end of the module, participants will have both a conceptual map of the multimodal-and-world-models landscape and the practical skills to train, fine-tune, and critically analyze such models on their own research problems.

The ultimate goal of data science is to drive effective decisions. While modern machine learning excels at identifying correlations, acting on these patterns can be misleading or even harmful. To make a real impact, we must understand causality—the 'why' behind the data—to know which levers to pull to achieve a desired outcome. This is the core challenge addressed by causal discovery.
This pursuit of causal understanding is foundational for the next generation of AI. It is the key to building genuinely explainable AI (XAI) that can justify its reasoning with causes, not just complex correlations. Furthermore, it is crucial for accelerating scientific progress, enabling researchers to unravel complex systems in fields ranging from medicine to economics.
Although identifying causal links traditionally requires experiments (interventions), this is often impossible, impractical, or unethical. The central challenge, therefore, is learning cause-and-effect from purely observational data. In this talk, after briefly surveying the field, I will discuss recent advances in this area, focusing on the fundamental problem of distinguishing cause from effect (i.e., does X→Y or Y→X?) from bivariate data.

In this tutorial, I will address the problem of the learning of deep models when data is biased. Data bias can be denoted as spurious correlations between data attributes and the labels. It is an important problem since we don't normally have the full control of data acquisition, so we cannot be sure that such data is collected properly, that is, representative of the true-world distribution, nor even we can be safe that bias is not still introduced even under strong human control, since humans typically bring own bias(es) and this is transferred to data as well. Bias can also originate from algorithmic choices and the employment of naive evaluation metrics, as standard optimization procedures often favor good average performance, disregarding the impact on less-represented sub-populations. This is particularly impactful when bias affects sensitive attributes, with more or less subtle fairness issues consequently araising. In this scenario, when a model is trained with biased data for a downstream task (e.g., classification), the performance is typically suboptimal because the model has gained a limited generalization capacity, so it may make errors when a sample is not affected by the bias. In other words, a biased model takes a decision on a test sample in dependence of the presence of the bias, not on the basis of the characteristic features of the object to classify.
In this context, there are typically two main cases, i.e., the bias can be known or unknown, leading to supervised and unsupervised bias mitigation approaches, respectively. In both cases, several approaches operate directly to unlearn that particular attribute, trying to alleviate its strength during the training procedure. These techniques include resampling, upweighting, regularization, or adversarial optimization. However, their application often requires a-priori information on the source of bias, or its estimation, which results in a key aspect of the bias mitigation algorithms. The unsupervised case is in fact more challenging, and surely more interesting and applicable to real scenarios, and several methods have been proposed to estimate the unknown bias, whose performance clearly depends on how well it is identified. Results on standard benchmarks and a few application use-cases will be illustrated to show that capabilities of the several approaches.

Machine learning systems are increasingly deployed in critical applications, where failures can affect security, privacy, trust, and consumer safety. However, these systems can be intentionally manipulated by adversaries who exploit weaknesses in the learning pipeline, the model behavior, or the interaction interface. In this seminars, I will introduce the foundations of machine learning security, discussing why security and safety must be considered core requirements in the design and deployment of modern AI systems. I will first present the main threats in adversarial machine learning, with a focus on evasion attacks, where carefully crafted inputs deceive a model at test time, and poisoning attacks, where an adversary manipulates training data to compromise the model behavior after deployment. I will then extend the discussion to emerging security and safety issues in large language models and foundation models, including prompt injection, jailbreaks, misuse, unsafe generation, and the difficulty of enforcing reliable alignment under adversarial interactions. Finally, I will discuss proactive measures for building more robust and resilient AI systems. These include systematic security testing, adversarial evaluation, robust training methodologies, and security-aware development pipelines. The goal of the seminars is to provide a unified perspective on machine learning security, connecting classical adversarial attacks with the new challenges introduced by large language models, and to outline open research directions toward trustworthy AI systems that are secure by design.

Generalization concerns the performance on new data, which is a core theoretical problem in machine learning. With the advent of deep models, generalization capability analysis has undergone fundamental changes, becoming one of the key issues in deep learning theory. This lecture will review the development of classical machine learning generalization theory, explain the significant challenges brought by deep learning, and introduce the relevant progress of our research group: i) functional space expansion, ii) dynamic low-dimensional subspaces, iii) sharpness-aware minimization. The future development of theoretical machine learning will also be discussed.